Al-Driven Precision Medicine/R&D Efforts in Academia and BioPharma
SESSION 9: AI-Driven Precision Medicine/R&D Efforts in Academia and BioPharma
Moderator & Introduction: Marco Gottardis (Gottardisbiotech LLC)
Generative AI for the Development of Novel Cancer Therapeutics
Alex Therien (Generate Biomedicines)
Picking Pockets in the Disordered Androgen Receptor N-Terminal
Andrew Allen (Peptone)
View the Transcript Below:
AI Driven Precision Medicine R&D Efforts in Academia and Biopharma
Marco Gottardis [00:00:07] Hello, everybody. Thank you for staying for the last session. I always appreciate you guys making the effort to stay to the end. I again, as every year, I want to thank Howard and Andrea for their collaboration on this session. And we look forward to many more, I hope. So, I was gonna have some comments, but I’m gonna shorten those because I’m gonna do it efficiently with this session. And I’m just gonna get right into the speakers because we’re already late, about 10 minutes. So now we’re gonna get into the drug discovery space. And the first speaker here is Alex Therien from Generate Biomedicines, and it’s gonna show you a new paradigm of what we think about wet labs and their relationship actually to computational biology and how we can move drug discovery faster. So, Alex? Oh, there you are. Okay.
Alex Therien [00:01:03] Just waiting in the wings.
Marco Gottardis [00:01:04] Yeah, there you go.
Alex Therien [00:01:06] Thanks to thanks to the organizers for the opportunity. Thank you all in the room, the handful of you that remained for sticking it out through this last session, especially because like me, you’re probably starving and looking forward to lunch. I have a bit of a confession before I start. I am not a prostate cancer expert. I am not an AI scientist, I’m not even a computational scientist. Don’t tell anyone, but I’m not an MD either. So, why am I here today at the Prostate Cancer Foundation Retreat presenting as part of an AI session? Well, hopefully I’m here to tell you a little bit more, teach you a little bit about Generate’s generative AI platform, and more specifically how we are translating it into promising novel protein biotherapeutics that we’re developing to treat multiple diseases across various therapeutic areas, including oncology, although not specifically prostate cancer, at least for now, at least as of yet. So perhaps a secondary objective of mine today is to inspire one or more of you in the room today to have some great ideas about how we could collaborate together to apply Generate’s platform to this critically important disease. These are my disclosures. I am a full-time employee of Generate Biomedicines. Perhaps more importantly, I am originally Canadian. I did my postdoc at the University of Toronto. I’m a huge Blue Jays fan. So, I know there’s a high potential for conflict of interest in this crowd, but go Jays, and I thought I had to mention that. For those of you who don’t know about generated biomedicines, I’m going to assume that’s most of you in this room. We are a clinical stage end-to-end drug discovery and development company that was founded on the hope and now the realization and the success of a generative AI platform aimed at designing and developing novel protein biotherapeutics. Now, if our CTO was here today, he might disagree with that particular definition. He might say that we are a tech company first and foremost, and I think that there is room for both of those opinions, but it’s hard for me to look at our pipeline and not think of generated biomedicines as a drug discovery and development company. You can see the pipeline here. Essentially in seven short years, we have put together a robust and expansive pipeline of protein biotherapeutics that include as a first wave multiple high-affinity, half-life extended monocle antibodies that we are developing for various diseases in the I & I space. And that includes GB0895, which is our lead molecule at the top there, which we’re developing for severe asthma and COPD, and is actually phase three ready, which I think I could be wrong, but I think, makes it the most, or at least one of the most advanced AI generated antibodies ever to be developed as a drug. So that’s exciting for us. So, in addition to our plethora of monospecific and bi-specific I & I monocle antibodies, you’ll see in the middle here a handful of oncology programs, four specifically, two of which we are codeveloping with our partners Roswell Park and MD Anderson. And so, I’ll spend about half of the remaining time telling you more about our platform and how we’re trying to use it, and then the rest of the time telling you more specifically actually showing some data around some of these oncology programs. So, while you are distracted by the animation on this slide here, I’ll tell you a little bit about our vision. It’s really divided into two parts on the left-hand side is really more of a short-term vision that we actually think we’ve essentially accomplished at this point. And on the right-hand side is something that is much more transformative, something that is much more aspirational for us. So, I think we have already demonstrated that we can create transformative medicines that are essentially out of reach of traditional technologies. Or at the very least that we can do it cheaper and better than companies that are using traditional methodologies. On the right hand side is something that we see happening over the next five to ten years perhaps and it really sort of encapsulates this idea that we want to be part of what I think, what we all think Generate is going to be a complete transformation of the way that we do discover and develop drugs in the next five to ten years. And it is going to be driven by advances in AI technologies. And so, we’re very excited about that. We want to be part of that. Our platform is very holistic. It is not just about AI, but at its very heart is our generative AI model, or actually I should say our generative AI models, because we have several of them. And so, I thought it’d be a good place to start to make sure that we’re on the same page about what generative AI is. And what better way to do that than to ask a generative AI platform such as ChatGPT? Yes, ChatGPT is a generative AI platform. It is of course different from our platform. Whereas ChatGPT was trained on something like, I think close to a petabyte of text data. And what it spits out when you give it a query is words, which hopefully are in the right order and actually explain the question that you asked it. We know that it doesn’t always do that, but it generally does. Our generative AI platform is trained on things like protein structure, protein sequences, protein function, binding affinities, and all sorts of other protein parameters. And what it generates when you give it a query are sequences of proteins that will hopefully fold in the structure that you’ve asked for and once it has folded have the function and the biophysical properties and any other properties that you’ve asked for it to generate. And again, just like ChatGPT it does a good job most of the time, but not all of the time. Just to dive into a little bit more around our vision and what we’re really trying to accomplish, we’re really trying to transform the current paradigm of drug discovery, which is laborious, high-cost, and if we’re honest with ourselves, often relies more on luck than on actual data. We’re trying to transform that into a, what we call, a programmable approach to drug discovery and development. We’re trying to essentially put in place a process that will allow us to generate proteins that can bind to essentially anything we can imagine. Generating molecules that have whatever functional characteristics, whatever binding characteristics, biophysical characteristics you want it to have. A process that will allow us to make drugs that are modality agnostic, so across multiple different kinds of proteins and in fact beyond proteins eventually. And allows us to prosecute multiple biological hypotheses at the same time. We can optimize across multiple parameters at the same time because of the scale not just of our dry lab efforts, but of our wet lab efforts as well. And then again, ultimately the goal to deliver the right drug for the right disease to the right patient much cheaper and much faster. So again, at the core of our platform is our generative AI models. What these models allow us to do is ask them questions like can you make me a protein that will have a particular structure, that will fold any specific shape, that will assemble with a particular symmetry, that will have a particular functional parameter. We published an early version of our AI model called Chroma back in 2023. And what you can see is that it does exactly that. If you ask it, for example, if you look at the top right of the slide there, if you ask it to generate proteins that will assemble in a very predictable way into large oligomers, it can do that. And you see various examples of that at the top right. You can even ask it to make proteins or to spit out a sequence of a protein that will fold as a letter of the alphabet or a number. And you can see examples of that on the lower right-hand side. You can also ask it to fold, make proteins that fold exactly like known proteins, proteins of known structure, but with a different sequence, and there are examples of that in the lower left-hand side. So, you can imagine, I think, the potential power of a platform like that if it gets applied to drug discovery. And in fact, in our hands, it has been extremely powerful. So, we have made antibodies, we have made beyond sort of the vanilla monospecific antibodies, we’ve made de novo antibodies. So that means essentially making antibodies that bind to any structure or shape that we feed into the model. We have made bi-specific antibodies, antibodies that have improved characteristics like affinity and potency, improved internalization, which is very relevant for ADCs, for example. We’ve also made enzymes for ERT, enzymes that are able to evade the immune system, so decreasing immunogenicity, which is critical for ERT. We can make membrane proteins and apply those to cell therapies. I’ll talk about that a little bit later on. We can make growth hormones and optimize them to bind to a very particular receptor and not another receptor of the same family, for example. And we can, of course, modulate the affinity as well, and we can make many proteins and peptides. So, again, there’s really no limit to what we’re able to do with this platform if it’s applied properly. But as I said earlier on, our platform is more than just AI. We try to push the limits of innovation both on the dry labs and on the wet lab side. What we essentially are aiming for is what we call the generate measure learn cycle, which you can see in the middle of this of this slide here. Essentially, we start with our generate AI ML models, we ask it to make proteins that have particular characteristics. We are able to then make these proteins using state-of-the-art DNA and protein synthesis technologies. We can then measure these proteins using a variety of again innovative assays. We use microfluidics, for example, to escape you know plate-based assays. We have a state-of-the-art cryo-EM facility, so we can generate high-resolution structural data, which is extremely powerful in an AI model like this. We have generated hundreds of structures already, and we’re hoping to get to about a thousand by the end of the year, which if anyone who knows cryo-EM knows this is a tremendously large number of structures to generate. And we push the limits of multiplexing and automation as well, so that we can ask multiple questions at the same time and try to optimize among multiple parameters as opposed to just a single parameter. So, what does this all mean? Does this actually allow us to address that vision that I mentioned off the top of completely revolutionizing the way that drugs are discovered and developed? Probably not on its own. But we do think that if AI is applied across the entire spectrum of discovery and development, that we can actually get to that vision. And I’ll go through this very quickly, but essentially, we know there are companies already using AI for target identification. We’ve seen examples of AI being used for patient stratification to predict how patients will respond to particular treatments. There are companies out there, and we know that this has this has actually worked, that AI can be used to design more efficient, faster clinical trials that will get you to your answers faster. And so, we think that all together these various applications of AI taken together will allow us to achieve that vision that I mentioned previously. So, I now want to switch gears and talk a little bit about how we’re using our platform at Generate to discover and develop drugs in the oncology space. And we think that our primary advantage in oncology is that we can basically make binders to any shape that we want. Specifically for oncology, if you have a tumor antigen specific or associated antigen that may not be amenable to generating antibodies in the traditional way. They may not be immunogenic, you may not be able to generate antibodies against it. Well, we know we can make an antibody as long as we know what the shape of that antigen is. And so, we can then convert these binders to a variety of immune-based therapeutics like ADCs, T-cell engagers, and CAR Ts. And we have four programs currently a generate that fall within those categories. We’re designing an ADC of our own where we’re trying to optimize for internalization to maximize the ability of the payload to kill the cancerous cell. Another ADC related program, which I’ll talk to you a little bit more as I show some data in a couple of slides, is around a neutralizing antibody that binds to free MMAE, which is a payload for some of these ADCs, like enfortumab vedotin, for example, that tend to sometimes fall off the ADC, go into circulation and hit tissues and cells that it’s not meant to hit, causing adverse events and toxicities that can actually cause some patients to ask to discontinue the drug even if it’s working for them. So that’s a really, really bad thing. And we’re developing a neutralizing antibody that will bind to that MMAE and prevent those AEs. We are working with MD Anderson to develop a CD3-based T-cell engager. Again, I can’t disclose the targets at this point, but hopefully by next year those things will be out in the public domain. We’re also developing a CAR T with Roswell Park. Okay, I promised to show you a little bit of data. I’ve got three slides of data and then I’ll be done. These are some data that we generated associated with our CAR T program. What you’re seeing here is that every one of those dots is essentially a single variant that was generated by our platform and that we expressed and tested for interferon gamma release, T cell proliferation, and ability to kill tumor cells. And what you can see is that in that very first cycle, we were able to generate a very large number of variants that have all of the right characteristics for a CAR T, namely high T-cell proliferation, high interferon gamma release, and high tumor killing. I can tell you that if we were to take these data, I can’t show you the data unfortunately, but I can tell you that if we were to take these data, feed it back into the model and ask it to spit up a next cycle of variants, it would do a lot better. You would see all of those dots migrating towards the upper right-hand quadrant of this graph, and they would all turn red. Typically we only need to do this once or twice to get to a development compound. This is a rather complex slide, but with a very simple message. We can make TCEs using our platform of various formats, you can see on the left-hand side of the slide. We can measure various functional endpoints of these TCEs. On the top right is T-cell activation. On the bottom right is cytotoxicity, where you can see in the green line killing cells that express the antigen, but not cells that don’t express the antigen. We feed those data back into our models again. We can optimize across multiple different parameters, and within one, two or maybe three cycles, we end up with a development compound like molecule that we can push towards the clinic. This is my final data slide showing some very recent data around our MMAE neutralizer. So again, the MMAE is a payload for some ADCs like in enfortumab vedotin. Patients that are treated with this ADC tend to have AEs that are associated with free MMAE in circulation. We have shown in mice that we can deplete up to 80% of the exposure to this free MMAE without impacting the efficacy of the ADC whatsoever. And the data that you’re seeing here are in nonhuman primates and homologous monkeys specifically, where we can show that at different doses, we decrease the exposure to free MMAE in the 50 to 80% range, and that that is associated with a decrease in both neutropenia and skin lesions that are often associated with use of these ADCs. So very excited about this program. This program is moving rapidly towards the clinic. I’m a minute over time. Rather than just bore you with recapping what I just told you. I will just perhaps bring back the words of our keynote speaker yesterday, Mike Milken, who exhorted us all to embrace AI and machine learning, I think it is inevitable and it should be clear to all of you that AI is going to transform the way that we discover and develop drugs in the future and we would be fools not to sort of embrace that and make it happen because it will allow us to bring the right drug for the right disease to the right patient much faster and cheaper and that’s good for patients and it’s good for society. And I will end here. And happy to take any questions.
Unknown [00:17:32] Really nice talk. Could you comment on the affinity range of your de novo binders?
Alex Therien [00:17:39] Yes. Typically, the first hits, so for de novo, typically for the first hits we would be lucky to get a low micromolar type of affinity. But what our platform is extremely good at once you have a binder is making it a lot better. So, we have made antibodies with within a few cycles we can get it down to the femtomolar range. So extremely high affinity.
[00:18:04] Thanks.
Marco Gottardis [00:18:06] I think we can ask Alex more questions after
Alex Therien [00:18:10] Yeah, sure. Absolutely. Thank you, everyone.
Marco Gottardis [00:18:21] I want to introduce my friend Andrew Allen who’s from Peptone and he’s gonna be talking about how we can use AI in terms of attacking disordered proteins like the androgen receptor. Please. Welcome.
Andrew Allen [00:18:40] Thank you, Marco. Thank you, everybody, for your patience. And thank you to Howard and Andrea for inviting me to be part of this session today. So, we’ve been in some very elegant high-level AI land, and I’m going to bring it home a little bit now and bring an application of the new to the old and talk about new ways to tackle the androgen receptor. We’ve been hearing about proteins, and of course, many proteins have had their structures identified originally through X-ray crystallography, and then more recently with newer techniques, including NMR and then cryo-EM most recently. Those structures have enabled massive amounts of training and the development of elegant, sophisticated tools recognized obviously by the Nobel Prize Committee, and we now have Alpha Fold, Boltz-2, and a variety of tools that can take from primary sequence to protein structure. However, many proteins have regions of disorder, so termed, where we do not have crystal structures. There is basically no training data, and so we have no idea about how to drug these proteins using these elegant newer techniques. Many proteins have regions of disorder. So even if there are parts which are druggable, which are crystallized, and we understand, there will be regions which are disordered. These are referred to as intrinsically disordered regions of proteins. These are the ones that we at Peptone are interested in. Now, when we refer to IDPs, intrinsically disordered proteins, it’s a little bit of a misnomer. In truth, these are ordered proteins, but they just have multiple forms of order. And so energetically, they don’t have a single low energy conformation, which the protein stably adopts. And instead, they have multiple conformations and move dynamically between those different structures. And that’s really what we’re referring to when we talk about IDPs. And the question is, of course, how can I discern structure of these disordered proteins? Peptone was founded a few years ago by a biophysicist, Kamil Tamiola, company’s based in southern Switzerland. And Kamil and team developed a novel approach to identifying experimentally the structure of disordered proteins. And it starts using a technique of hydrogen deuterium exchange mass spectrometry. And so, the notion is you’ll have a protein, and the protein may have a pocket in it. And this may be one of many conformations of that protein. And I want to label the pocket. And how can I do that in a way that does not disrupt the structure of the protein? How can I label in an elegant and minimally disruptive manner? And the simplest thing to do is to replace protons with deuterium. No change in charge and a very, very minor change in mass. And so, in principle, If I perfuse the protein with deuterium, I will now start to see dynamics within the protein. Superficial regions of the protein or shallow pockets will rapidly exchange deuterium for protons, and deeper pockets, clefts in the protein that exchange will take longer. So, the on and the off rates will be different. So, in principle, I can use hydrogen deuterium exchange mass spec to understand the surface topology of these proteins. The problem is that using standard techniques today, we do not have adequate resolution. The first time point that’s typically collected is at around one second, and after one second, all of the action has already happened. And so, you really gather no useful information. The innovation of Peptone is to do ultra-fast hydrogen deuterium mass spec, and in 10 millisecond increments, we probe the protein structure using deuterium using stop flow experiments. And you can see that now with this much higher resolution, you start to see evidence that there are pockets in these proteins, and these are evident here in this heat map in as the blue fingers projecting into the protein. So, what do I do with this? On its own, of course, this does not give me adequate resolution to know that much about the protein, and certainly not enough to launch a drug discovery campaign and to develop understandings about structure activity relationship or SAR, which is the mantra for most medicinal chemists. But what we can do is we can take our experimental data and combine it with the programs like Alpha Fold 3, Boltz-2, and so on. And we do this using obviously GPUs that have been developed and made available now by NVIDIA and others. We have a collaboration with NVIDIA, so we don’t actually buy huge numbers of GPUs, we basically rent them. And what we can do now is use Alpha Fold 3 or Boltz-2 or a derivative thereof to simulate billions of possible protein structures for these disordered proteins. But now I can compare those experimentally synthetically derived structures with our experimental data. And I can ask the question do any of those structures fit with the experimental data? And if they do, I will return that structure and I will then iterate. And by so doing, we start to figure out what the druggable pockets in these disordered proteins might be. So again, we’re still in the prediction world, but we’re using experimental data to inform our predictions. Now, when we compare the predictive accuracy of our system against these standards at high protein structure levels on the left here, obviously we have no reason to believe we will outperform any of the other techniques. But as the protein becomes more disordered as you shift to the right here, our predictive accuracy increases substantially above the standard techniques because there is so little training data. But how can we make drugs out of this? Well, what we’ll do is take our putative pockets in our proteins of interest, and we will do a virtual drug screen. So, we take a standard small molecule library of very simple small molecules from a standard CDMO, and we will screen those in silico against the modeled pockets. And if we identify any predicted binders, we simply buy those molecules. We’ll buy a few hundred, they ship them from the CDMO, it takes a couple of weeks, and then we run the HDX-MS experiment again. And so, on the left here, you can see what happens. Each row is a drug, and what we’re looking at is change in the HDX-MS data, where blue is protection. So, if the molecule binds as predicted to the pocket, you will prevent deuterium exchange. That’s referred to as a protection factor, and that’s manifesting here as blue. And you can see that obviously not all of our predictions are accurate, but many of them do indeed bind as intended into the pocket. And this enables us to simulate drugs and then experimentally confirm that they do indeed bind to the pockets. This whole process takes about four to six weeks. So, within a month and a half, you go from knowing nothing to developing putative pockets, identifying putative pockets, an in silico virtual drug screen, confirmation of binding, and then of course I’ll take some of those binders into functional assays and ask the question great, I’m binding to the protein. Do I now change function of the protein in a therapeutically useful way? So, how do we apply this technology to targets we all care about? Because what we’re doing is novel, we wanted to try and reduce target risk, so we didn’t pick new targets, we picked an old target, androgen receptor. And of course, you all know the AR extremely well. This is the classic four-domain structure of the AR, a member of this large family of nuclear hormone receptors, which includes some other very important proteins, ER, GR, MR, PR, etc. And classically, you have at the C terminus end the very well-structured ligand binding domain where DHT binds. That’s the little green structure you can see there. That’s DHT binding into the pocket. Then we have the hinge domain, thought to be largely responsible for dimerization, the DNA binding domain, and then the large N terminal domain, which is where the action happens, and the AR dimeric form that has been activated by DHT binding relocates to the nucleus, and then the N terminal domain recruits the co-transcriptional activators that we’ve been hearing about. And Dr. Chinnaiyan gave us a nice talk earlier, talking about some of the components of this complex. And we heard about that yesterday as well with regard to NSD2. So, drugging the N-terminal domain is an attractive prospect because, of course, we all know about resistance, drug resistant variants. We know that splice variants are very common and are associated with resistance to ligand binding domain inhibitors. And in theory, an NTD inhibitor would inhibit the function of those splice variants. If we make it small enough, maybe we can deliver enough drug into the nucleus to also outcompete the heavily amplified AR that we see in the typical drug-resistant patient. So, the desire for a small molecule NTD inhibitor has been around for a while. It’s just been a hard thing to solve because it’s disordered. And again, it means you have no idea of structure. And the best you can do are phenotypic screens groping in the dark for things that might bind. And companies have tried that. Most recently, Essa Pharma gave it a valiant effort, but the drug just not probably potent enough or drug-like enough to be active in clinical trials, as we all know. And I’ll show some data of their drug, masofaniten, in a minute, which does bind to the n-terminal domain, but just with very low potency. So, this is an order or disorder plot, and you can see the problem that the entire n-terminal domain is disordered. So, we applied our technology to this. We expressed full-length AR or just the N-terminal domain. We use HDX-MS to identify putative pockets, conduct the virtual screen, bring in those binders, and then screen them and confirm binding using HDX-MS again. And we ended up with a couple of series, and the PEP 335 series is the one that we’ve been driving now towards an IND. So, a little bit of medicinal chemistry. Each dot on this plot is a drug or a molecule. On the X-axis, we have log D, which is a proxy for permeability. You want to be over on the left-hand side, and you can see the challenge with Masofaniten, which had low permeability. And you want to be potent, and this is potency measure just in a full-length LN cap model, more potent higher up the y-axis. So, you want to be driving to the top left, which is obviously where we’re heading. And then, of course, we want to test against V-7 based systems, and so we have a couple of V-7 reporter systems. And again, you want to drive to top right. Obviously, the first generation utamides do not really work in this system because the LBD is gone. Masofaniten again, low potency, and you want to be driving up and to the right. So, we have molecules that we can then test in more resistant cell lines. Here we’re looking at a VCaP line. And in vitro on the left, we’re comparing to the LBD binders. And they do have activity here, of course. There is AR signaling in VCaP. But as you can see, the potency is relatively low, and the maximal effect is also pretty modest, maybe a halving of the proliferation. And you can see that our series are more potent, they’re left shifted, and they have a more complete effect, bringing proliferation down to the origin. And then when we compare with PROTACs, which obviously have been a very interesting class, maybe being displaced now by RIPTACs, but you can see that we have a pretty similar profile in vitro compared to the PROTACs, which of course in vitro behaved very impressively. We want to see suppression of AR inducible genes. And so again, when we’re looking in vitro, we see this. And on this plot, we’re just looking at four well known AR inducible genes. The greens are the first generation utamides, the gray is a reference PROTAC, and then the blues are dose escalation of one of our series, and you can see very nice, very complete inhibition of expression of a variety of genes here. Okay, now let’s confirmed that our molecule does indeed bind to the AR, and there’s a variety of ways of doing that. This is the HDX-MS plot. So, you can see the Tau5 region is where our drug seems to bind principally, but perhaps not exclusively. There is folding of these proteins, and so you’ll often see multiple points of protection, which is telling us something about protein folding, and there’s a lot of information that we’re learning from in experiments like this, teaching us something about how the protein folds in the presence or the absence of our drug. And it’s well known in the disorder protein world that your drug probably changes the nature of folding. But you can see we get nice protection in the Tau5 domain, and there’s a nice positive control. The DNA binding domain at the top there is protected by DNA in this assay system, as you’d expect. We can confirm binding by NMR, that’s shown on the right, which you can do once you’ve got a potent inhibitor. It just doesn’t work as a screening tool with low potency molecules. We can do competition assays. With synthetic androgen receptor ligands and you can see here that our series just does not compete with synthetic androgens. It just doesn’t bind in the same place. We’ve recently developed a nanoBIT assay which looks at dimerization. And here we’re using enzalutamide as a reference tool. So, on the left you can see that dimerization of full-length androgen receptor is nicely inhibited by enzalutamide as you’d expect, and that’s shown in the blue. But over on the right, when you’d use it in a V-7 system, we don’t see dimerization with enzalutamide, sorry, inhibition of dimerization with enzalutamide, but we do see it with our PEP series. And again, it’s not that we think we’re necessarily a direct dimerization inhibitor, but maybe just changing the conformation of the protein to prevent dimerization. And finally, we’re now moving into animal models and seeing nice growth inhibition. This is in an LNCaP model, and we’re right now moving into treatment resistant cell lines for in vivo testing. But obviously, we have molecules now that have oral bioavailability and adequate half-life. Our goal is to have a development candidate nominated in Q1 of next year and file the IND by year end, and off we go. And then finally, in my last minute, C-MYC was discussed this morning by Dr. Chinnaiyan and has been cropping up over the last couple of days. Dr. Brown discussed it as well, or L-MYC the other day. C-MYC, of course, is another compelling disordered protein. It has a region of structure. You can see those two alpha helices there. And there in the C terminal domain, they dimerize with MACs. And so, it’s this MYC-MACs interaction, which has classically been targeted by everybody trying to develop small molecule inhibitors because that’s the piece where we have insights on structure. But the MBII domain has been implicated in cancer formation. There are some pediatric malignancies, leukemias that have mutations in the MPII domain of C-MYC. And this interacts with the large scaffolding protein called TRRAP. And so, the interesting question has been: can we develop MBII binders which disrupt the MYC-TRRAP interaction? And the answer so far appears to be yes. So, using the same technology, we’ve developed a series of molecules that are now quite potent. We’re operating now at about 150 nanomolar inhibitors here, which appear to be inhibiting growth in a MYC-dependent fashion. And these are not non-specific inhibitors. And the biophysics is confirming the binding exactly where intended. So, in summary, ultra-fast hydrogen deuterium exchange mass spec appears to be giving us a way to develop small molecule inhibitors of disordered proteins. The AR is a very attractive and classic target. Of course, we’re interested in others like C-MYC. But of course, you could also use these binders as the basis for other classes of molecules, degraders, a hook for isotopes, or even for cytotoxins. And with that, I’ll close. I’d like to thank Dr. Aggarwal, de Bono, Chinnaiyan, and Alimonti, with whom we’re at various stages of collaboration. Of course, thank the Prostate Cancer Foundation, Mark Mansour is our collaborator on C-MYC, and the Peptone team as well. Of course, thank you for your attention.
Marco Gottardis [00:35:31] We can take a couple of questions.
Unknown [00:35:36] Hi, excellent talk. I’m wondering do you know what is the mechanism of action of your AR and terminal binder? Is it similar to the ESSA compounds inhibiting downstream transcription?
Andrew Allen [00:35:50] Yeah, great question. The simple answer is we don’t know yet. We’re figuring out the biology. As you can see, this is a chemistry led program, and then the biology plays catch up. Obviously, we know that we’re inhibiting AR transcription induction, we’re blocking AR dependent cell line proliferation. So, it’s doing the right thing, but the great question you’re asking is exactly how, and we’re figuring that out right now. We don’t know yet.
Unknown [00:36:14] And I have a second very quick question. We have identified the phenotypic screens AR and terminal domain binder that degrades the proteins and the variants. If we use our drug, could you identify with a deuterium exchange method the exact binding pocket?
Andrew Allen [00:36:33] Yes, simply put. Yeah, your presentation was very interesting. I love the work that you showed. This does have an effect on the amount of AR, but it we don’t think it’s acting as a direct degrader. There’s a lot of, I suspect, you know, consequences to disrupting the folding of the protein that probably changes the way in which it’s handled inside the cell and may indeed target it to the proteasome through sort of classical fashion versus the iatrogenic form that you’ve developed.
Unknown [00:36:58] Thank You.
Russell Szmulewitz, MD [00:37:00] Russell Szmulewitz, University of Chicago. That was great. Super cool. I can almost understand it, which is a credit to you. I have a question about how you screen for binding of other, you know, in this case, nuclear hormone receptor members or you know, other proteins in general. Do you have, you know, either an AI algorithm or do you have you know some sort of other thing that you do to say, hey, this isn’t just gonna bind a whole bunch of things and cause a lot of toxicity?
Andrew Allen [00:37:32] Yeah, that’s a really important question. So, the crude, simple approach that we take is to do in vitro screens with recombinant proteins to make sure we don’t inhibit them. But what we find is that we have molecules that inhibit the AR very nicely and don’t touch ER and GR and PR and so on. And then we have molecules that actually are pan inhibitors. So now the really interesting question is, well, what do they do? Where do they bind? What’s their function? And we don’t know. So, what’s interesting here is we’re generating now quite large quantities of data with all of this binding data. And of course, now the key is you’ve got to train from that. I think this theme came up yesterday that you don’t use your AI, you know, develop it once and then sort of apply it and never learn again. You apply it and then you keep learning. You have to generate large amounts of data in order to keep learning. You know, you can’t learn off of two molecules or something. But if you’re doing it at scale with hundreds and thousands, then you start to learn. But we haven’t done HDX-MS systematically across the entire field of disorder proteins, right? And that’s an interesting thing that will come over time, but we’re obviously a long way from that currently.
Bob Stein [00:38:38] Hi. I’m Bob Stein. I enjoyed your talk. Have you been able to demonstrate that your small molecules prevent interaction with coactivators?
Andrew Allen [00:38:47] No, the that’s the question I’ve been pinging over to my colleagues over the last three days. Like, we need to get on top of this question. What are we doing specifically for the recruitment of those transactivators? And of course, maybe we could think about building a program to specifically disrupt some of the oncogenic forms. Obviously, the work from doctors Chinnaiyan and Chen suggests, you know, could we disrupt AR-NSD2 interaction in a selective fashion? Because a catalytic NSD2 inhibitor is likely to have a whole bunch of toxicity, of course, because that’ll systematically or systemically I should say disrupt NSD2 function, which may be bad for us. We don’t know yet, really. We may learn from that k36 company that we heard about. But I think could we do it in a selective fashion, so we only disrupt that oncogenic interaction and leave you know normal physiological function of NSD2 undisrupted? That would be very compelling. Yes, great question. Thanks.
Bob Stein [00:39:42] Thank you.
Nada Lallous, PhD. [00:39:43] Nada Lallous, University of British Columbia. I’m wondering because also IDR has been shown to be involved in phase separation and I guess those pocket could also have some stickers like hydrophobic residue. Did you look into that? Did you integrate that into your model and for example the NRs that the compounds that bind same NR and blocks that, are they same areas and could they be involved in phase separation, for example?
Andrew Allen [00:40:12] Yeah, it’s a very good question. We have not looked yet. As I say, the biology is catching up. So, if anybody’s interested in collaborating, we’d be happy to talk to you about figuring this out. ‘Cause we’re a small company our biology is lighter than our chemistry. Thank you very much.